When refining margins tighten, every basis point recovered from operations reaches the bottom line. The difference between outperforming peers and treading water is then set less by capital projects and more by day-to-day decisions inside the plant. A small temperature drift between units, a delayed response to a crack-spread swing or a single unplanned outage can each erase margin the plan counted on.
Four common issues account for much of that loss: unit-to-unit variability, slow response to market swings, reactive maintenance and the limits of control room bandwidth. Each can be addressed with the equipment and data the refinery already has.
Unit-to-unit variability
On a typical day the control room shows it: a one-degree drift in reactor temperature on one unit, a small pressure swing on another. Each looks harmless. Across interconnected units such as the FCC, hydrocracker and reformer, they add up to lost yield.
Process variability covers the swings in performance that come from changing feedstock properties, aging equipment, process upsets and well-intentioned operator adjustments. The cost shows up in several places:
- product giveaway, or sales into lower-value channels
- lost throughput when one unit throttles another
- extra energy use, and labor hours spent on constant adjustment
- downgraded product, which in a tight market can wipe out a week's margin gain
Conventional advanced process control (APC) was built to steady individual units, and it does that well. It was not built for the nonlinear interactions between units. Its linear models also lose accuracy when crude assays shift or catalysts age, which keeps engineers in a cycle of re-tuning.
A closed-loop optimization layer can treat the refinery as one connected system. A model learned from the plant's operating history, with first-principles constraints, captures the cross-unit relationships, and a controller trained by reinforcement learning (RL) writes setpoints through the existing APC to the distributed control system (DCS) as conditions change.
Slow response to market swings
Crack spreads can move by the hour, and margin moves with them. Many planning teams still rely on linear program (LP) models that refresh daily or weekly. That cadence made sense when product demand shifted slowly. With uneven demand and shifting crude differentials, it leaves front-line operations steering by an old signal.
The limitation is structural. An LP optimizes a snapshot, and planners then translate its targets into operating setpoints by hand. By the time those moves propagate through scheduling, the market has moved. An intraday price move can make the morning's optimal slate unprofitable by midday while operators are still steering toward the old plan. The losses compound: tanks fill with the wrong blends, demurrage costs rise, and valuable feedstock ends up in low-value products.
When the prices behind the optimization, such as product values and regional crude differentials, are updated through the day, the optimizer can recalculate netbacks as the market moves. The updated economics then move the targets that matter in the short term, such as cut points and blending ratios, and give planners a current signal for crude selection. None of this requires new equipment.
Fast market response is of limited value, though, if production is interrupted by outages that maintenance could have prevented.
Reactive maintenance and unplanned downtime
An electrical fault, a weather-related power dip or a seized pump can back up the whole crude slate. Each surprise shutdown puts the site into firefighting mode, reacting to alarms after the damage is done.
The cost is large. A major unit sitting idle loses production every hour, and emergency repairs add equipment rental, rushed contractor fees and knock-on disruption. Avoiding a single unplanned outage on an FCC or hydrocracker protects both throughput and cash flow.
Maintenance practice tends to swing between two extremes: conservative preventive work that takes equipment offline too early, and reactive repairs that come too late. Weak coordination between operations, maintenance and planning stretches outages, and partial inspections can miss the corrosion or fatigue that triggers the next shutdown.
Predictive models offer a better path. Trained on years of historian data the site already collects, they learn the drift patterns that precede failure, such as temperature creep in a reactor or rising vibration in a compressor, and alert the team well before the trend reaches an alarm limit. They can recommend the smallest practical maintenance window and fit it into an existing turnaround, without new sensors or equipment. Because the models are updated as conditions change, reliability work keeps pace with the plant.
Control room bandwidth
At shift change the board shows thousands of tags, alarms sound every few seconds, and handwritten logs wait to be read before anyone touches a setpoint. With that much data and competing KPIs, the safest move is often to leave temperatures or hydrogen rates where they are. That protects quality, but it gives away energy and yield in exchange for stability.
The problem grows when planning, blending and maintenance each optimize their own part of the plant in isolation. Without a shared, current picture, short-term schedules drift from what the plant is doing, and blending decisions waste valuable components or force reprocessing.
A recommendation layer helps close that gap. Models trained on historian data and lab results track the whole unit at once and recommend specific moves, such as lowering a furnace outlet temperature by one degree, that move the whole system toward better margin. Operators decide whether to accept or postpone each move, and their responses become part of the record the model learns from. Because the reasoning is visible, operators can see why a move matters, and the same models give newer staff a simulator to see the result of a change before they make it. This advisory stage is also where crews build the trust to let the controller act in closed loop within limits they set.
Running closer to the optimum also cuts excess fuel use and CO₂ emissions, and frees hydrogen or steam for higher-value uses, again without new capital.
How Imubit applies this
In refining, Imubit applies Map, Model and Control to these losses. Map traces each margin leak to the recurring decisions behind it, such as reactor severity, cut points and blend targets. Model uses Deep Learning Process Control® (DLPC®) to learn each unit's nonlinear dynamics from its own history, guided by first principles. Control takes those decisions to closed loop, writing setpoints to the existing APC and DCS minute by minute, inside limits set by plant engineers. In refining, deployments have delivered a $0.25/bbl margin improvement, and Imubit is trusted by 7 of the top 10 U.S. refiners. Read more about refining.
